Visual Verification

You verify that UI-facing changes render correctly through a screenshot-analyze-verify loop. This skill was extracted from runtime-verification so the visual workflow has its own home; the two skills compose (runtime-verification handles build/server/smoke/E2E/LSP-diagnostics; visual-verification handles browser-rendered UI).

Iron Law

UI CHANGES MUST BE VISUALLY VERIFIED OR EXPLICITLY SKIPPED. A build that succeeds and a test suite that passes do not prove the page actually renders. Every UI-relevant change goes through this loop or surfaces a structured SKIP/BLOCKED result that the completion gate can reason about — never a silent skip.

The visualVerification.requireVisualVerification setting controls escalation behavior, not whether the loop runs. Even when requireVisualVerification: false, an unattempted UI change must produce SKIP_WARN with the reason — the completion gate emits the warning so the user knows visual verification was not attempted.

UI Relevance Detection

Visual verification activates when EITHER signal fires:

# Signal 1: UI file extensions in the diff
git diff --name-only HEAD~1..HEAD | grep -iE '\.(tsx|jsx|vue|html|css|scss|svelte)$'

# Signal 2: UI keywords in the acceptance criteria
# (The invoking command passes the criterion list as input; check for any of these tokens.)
# Tokens: UI, page, display, render, visual, layout, responsive, component, style

If neither signal fires → emit SKIP — no UI-relevant changes detected. and exit. This is a legitimate skip, distinct from SKIP_WARN.

Browser Tool Discovery (priority cascade)

Try these in order; use the first available:

1. Playwright MCP (browser_navigate, browser_take_screenshot, browser_console_logs) — full capability: navigation, screenshots, console logs, DOM inspection
2. Chrome DevTools MCP — screenshot + console + DOM inspection
3. CLI fallback: npx playwright screenshot http://localhost:$PORT/ $SCREENSHOT_DIR/page.png
4. External skill fallback: Skill(compound-engineering:test-browser) — if the compound-engineering plugin is installed
5. External skill fallback: Skill(compound-engineering:agent-browser) — if the compound-engineering plugin is installed
6. No tools available → return SKIP_WARN or BLOCKED based on visualVerification.requireVisualVerification (default: false → SKIP_WARN; explicit true → BLOCKED, command-level escalation needed)

The skill does NOT silently install Playwright. Installation is a side effect with footprint; the user is asked via the command-level escalation path when the cascade falls all the way through.

Screenshot-Analyze-Verify Loop

Bounded by settings.json → visualVerification.maxIterations (default: 3). Iterate up to the cap when fixes are applied between rounds; halt earlier when verification passes.

For each page URL (dev server root + key pages from routes):
  1. Navigate to page URL
  2. Take screenshot → save to $SCREENSHOT_DIR/{page}-{viewport}-{timestamp}.png
  3. Read screenshot with the Read tool (Claude analyzes visually)
  4. Classify findings using the canonical schema in references/finding-schema.md (category=visual):
     - Blank page → P1 (blocks completion)
     - Render-blocking console errors → P1
     - Layout breaks / broken grid → P2
     - Missing content that should be visible → P2
     - Minor styling issues → P3
  5. If MCP tools are available: also fetch browser_console_logs and grep for JS errors / React warnings / CSP violations
  6. Record screenshot path as evidence (referenced from the per-criterion evidence bundle)

Findings emit using the canonical two-column Finding | Suggested Fix table (see references/finding-schema.md): bold {ID} · {category} · {location}` on the first line, problem prose after a <br>. Use the INT- prefix when invoked from integration-verifier, the VIS- prefix when invoked standalone. Location for visual findings is the URL path (e.g., http://localhost:3000/login instead of file:line`) — the schema accepts non-file locations for renderer-surface findings.

Responsive Verification

For each viewport in settings.json → visualVerification.viewports, resize the browser and repeat the screenshot-analyze step:

• Default viewports: Desktop (1280×720), Tablet (768×1024), Mobile (375×812)
• Per-viewport checks: content cut off, navigation broken at breakpoint, horizontal scroll on mobile, fixed-width elements overflowing the viewport
• Each viewport is a separate finding source — a layout that works on desktop and breaks on mobile produces a P2 with location like http://localhost:3000/ @ mobile (375×812)

Task Tracking

Create the visual-verification task suite upfront and update status as the loop progresses:

# Setup — create all visual verification tasks upfront
TaskCreate("Visual verification", "Screenshot-analyze-verify for UI-facing changes")
TaskCreate("Browser tool discovery", "Detect available browser automation (Playwright MCP, Chrome DevTools, CLI)")
TaskCreate("Responsive check", "Verify UI across configured viewports (desktop, tablet, mobile)")

# Browser tool discovery
TaskUpdate(browserToolTaskId, status: "in_progress")
# ... detect tools ...
TaskUpdate(browserToolTaskId, status: "completed", result: "{tool} detected")

# If not applicable (no UI files, no UI criteria) — legitimate skip:
TaskUpdate(visualVerificationTaskId, status: "completed", result: "SKIP — no UI-relevant changes")
TaskUpdate(responsiveTaskId, status: "completed", result: "SKIP")

# If no browser tools found — check requireVisualVerification setting:
#   requireVisualVerification: false (default) → SKIP_WARN
TaskUpdate(visualVerificationTaskId, status: "completed", result: "SKIP_WARN — no browser tools. Install Playwright MCP or use /flow:setup.")
TaskUpdate(responsiveTaskId, status: "completed", result: "SKIP_WARN — no browser tools")

#   requireVisualVerification: true → BLOCKED (command-level escalation)
TaskUpdate(visualVerificationTaskId, status: "completed", result: "BLOCKED — requireVisualVerification is true but no browser tools available")
TaskUpdate(responsiveTaskId, status: "completed", result: "BLOCKED")

# Screenshot-analyze-verify loop (when tools ARE available)
TaskUpdate(visualVerificationTaskId, status: "in_progress")
# ... for each page: screenshot → analyze → record findings ...
TaskUpdate(visualVerificationTaskId, status: "completed", result: "PASS/FAIL — {pages} checked, P1:{n} P2:{n} P3:{n}")

# Responsive verification
TaskUpdate(responsiveTaskId, status: "in_progress")
# ... for each viewport: resize → screenshot → analyze ...
TaskUpdate(responsiveTaskId, status: "completed", result: "PASS/FAIL — {viewports} tested, findings: {summary}")

Use TaskList after all visual verification completes to confirm all sub-tasks resolved.

Result Vocabulary

Output Format

### Visual Verification

| Check | Status | Details |
|---|---|---|
| Browser tools | {tool name or NONE} | Cascade result |
| Visual check | PASS/FAIL/SKIP/SKIP_WARN/SKIP_USER_APPROVED/MANUAL/BLOCKED | {pages checked, findings} |
| Responsive | PASS/FAIL/SKIP/SKIP_WARN/MANUAL | {viewports tested} |
| Console errors | PASS/FAIL/SKIP | {error count} |

### Visual Evidence
| Page | Viewport | Screenshot | Status | Findings |
|---|---|---|---|---|

### Visual Findings (canonical finding-schema)

#### P1 — Critical (Blocks Completion)
| Finding | Suggested Fix |
|---------|---------------|

#### P2 — Should Fix
| Finding | Suggested Fix |
|---------|---------------|

#### P3 — Consider
| Finding | Suggested Fix |
|---------|---------------|

Active Problem Solving

Do NOT silently skip visual verification. Actively solve tooling problems:

External dependency: compound-engineering (optional, documented)

The browser-tool cascade includes two entries from the compound-engineering plugin (compound-engineering:test-browser, compound-engineering:agent-browser) as fallback positions 4 and 5. Decision: document, do not vendor. Rationale:

• The cascade gracefully handles the absence of these external skills. Playwright MCP, Chrome DevTools MCP, and the CLI fallback (npx playwright screenshot) are all viable for the production loop; the external skills exist as additional fallbacks for environments where neither Playwright MCP nor a working npx setup is available.
• Vendoring the logic from compound-engineering would fork the behavior — the upstream plugin may evolve its browser-tool surface, and a vendored copy would silently diverge. The flow plugin would inherit maintenance for code it did not author.
• The compound-engineering plugin is part of the same Synapti marketplace and is recommended for installation alongside flow when visual verification is critical to the workflow. The cascade documentation makes the dependency relationship explicit so users in tooling-constrained environments can install it as a remediation step rather than discovering the absence at verification time.

What this means in practice

• Without `compound-engineering` installed: visual verification works through positions 1-3 of the cascade (Playwright MCP, Chrome DevTools MCP, CLI fallback). Positions 4-5 are skipped silently.
• With `compound-engineering` installed: positions 4-5 become available as fallbacks if positions 1-3 are unavailable.
• `requireVisualVerification: true` configured but no browser tools at all (none of positions 1-5 succeed): the skill returns BLOCKED, which the consumer escalates per references/escalation-format.md with three options (skip with note, manual verification by user, install browser tools).

The README documents compound-engineering as a recommended companion plugin in the Hook Compatibility / Dependencies section.

Integration with runtime-verification

runtime-verification is the parent verification skill — it owns build, dev server, smoke tests, E2E, LSP diagnostics, and acceptance-criteria verification. When the diff is UI-relevant, the consumer (commands/start.md Phase 4 step 8 / commands/pr.md Phase 4 step 4) invokes BOTH skills:

• Skill(runtime-verification) for build + server + smoke + E2E + LSP diagnostics
• Skill(visual-verification) for the browser-rendered UI loop

The two run in parallel when the dev server is up; if the dev server fails to start (a runtime-verification failure), visual-verification returns SKIP with the reason "dev server unavailable". The completion gate treats the dev-server failure as the primary finding.